Researchers at the Korea Advanced Institute of Science and Technology (KAIST), Jeonghyeok Do and Moonchurl Kim, have built an artificial intelligence model called GeoCore-9B that can reconstruct parts of Earth's surface obscured by clouds in images taken from space. Earth.com reports on the development, according to infohub.kz.

The new system addresses the main weakness of optical satellite imaging. Ordinary cameras capture reflected sunlight, so they are useless in overcast weather or at night, which is especially problematic when assessing floods and wildfires. A preprint of the study has been accepted for presentation at the NeurIPS 2026 conference.

Unlike earlier counterparts, which were retrained from ordinary photographs, the new model was trained from scratch exclusively on 10 million satellite frames with text descriptions. Each image was accompanied by precise coordinates and scale so the algorithm could understand the real size of objects. During training, a second neural network acted as a mentor, monitoring the correct joining of roads and field boundaries.

In tests, GeoCore-9B outperformed specialized cloud-removal programs in the accuracy of color and terrain structure reproduction. The algorithm also learned to convert data from radar satellites, which can see through clouds, into conventional color images.

The developers caution that the generated picture is a reconstruction rather than a real photograph, so the technology could potentially be used to create fakes about disasters or military events.